论文纸片盒子 07
Machine Learning for Fluid Mechanics
abstract
- 分析了流体力学领域中机器学习应用的历史、现状、存在的机遇。
- 归纳了机器学习的主要方法及其优势和局限性。
- 机器学习方法的引入，扩宽了流体力学研究的边界以及工业应用。
introduction
historical overview
challenges and opportunities for machine learning in fluid dynamics
- 流体力学与图像识别、广告推荐等不同，流体力学领域重视对于具体现象的潜在物理机制的量化分析。
- 流体力学领域问题通常具有非线性、多时空尺度的特征。
- 流体力学领域试验通常难以重复，模拟通常需要大型计算机长时间运行。
machine learning fundamentals
image
supervised learning
neural networks
classification: support vector machines and random forests
unsupervised learning
dimensionality reduction I: POD, PCA and auto-encoders
dimensionality reduction II: Discrete principal curves and self-organizing maps
clustering and vector quantization
Semi-supervised learning
generative adaptive adversarial networks(GAN)
reinforcement learning
stochastic optimization: a learning algorithms perspective
important topic we have not covered: bayesian inference, gaussian processes
flow modeling with machine learning
- 机器学习在流体力学领域的两个方向：1. 维度降低（dimensionality reduction）；2. 降阶建模（reduced-order modeling）
- 流体力学领域机器学习建模分类：1. 通过提取流体特征建立运动学模型；2. 通过利用各种学习框架建立动力学模型
flow feature extraction
模式识别和数据挖掘是机器学习的核心优势。
dimensionality reduction: linear and nonliear embeddings
clustering and classification
sparse and randomized methods
super resolution and flow cleansing
modeling flow dynamics
计算模型的关键目标之一在于平衡效率和准确性。
linear models through nonliear embeddings: DMD and Koopman analysis
neural network modeling
parsimonious nonliear models
closure models with machine learning
challenges of machine learning for dynamical systems
flow optimization and control using machine learning
stochastic flow optimization: learning probability distributions
flow control with machine learning
neural networks for control
genetic algorithms for control
flow control via reinforcement learning
image
